The Evolution Of Social Networks Driven By User Behaviors

The great value of social networks has been widely demonstrated, studied and utilized in both the industria and academia. Many upper-layer applications rely on the exact analysis of the social network structures, which are always evolving with time. Therefore,it it an important fundamental task to infer the evolution process, analyze the evolving structure and further predict evolution trend. Unlike traditional complex networks, the social networks are composed by people, and the evolution of social networks is driven by user behaviors. In this dissertation, we study the social network evolution inference and prediction problem based on user behavior modeling. The primary contributions are summarized as follows.1. We address the tie direction inference problem in undirected social networks. We discover four interesting directionality patterns in large-scale real-world directed social networks. Based on that, we design the general ReDirect optimization framework targeting on minimizing the pattern inconsistency of a directed network, and proposed e?ective algorithms for tie direction inference upon ReDirect ReDirect can be used in a general preprocess manner to recover the latent directions of undirected social networks and benefit other network analysis tasks.2. We address the network evolution process inference problem. We propose a new structure, LaFT-Tree, to express the network evolution process based on the transitivity of friendship. LaFT-Tree provides a hierarchical view of the flat structure of one’s egocentric network by visualizing its expansion trace. We design a probabilistic generative model LaFT-LDA for depicting the friend-making behaviors. The inference algorithm for LaFT-Tree based on the cascade LaFT-LDA inference is proposed.3. We address the social link prediction problem. Upon LaFT-LDA, we propose an interest-aware behavior generative model LFPM and its inference algorithm based on Gibbs sampling. With the inferred LFPM models, we build the latent friendship propagation network(LFPN) which depicts both the historical traces and future trends of latent friendship propagations in the networks. We propose a behavior-driven link prediction algorithm, LFPN-RW, which models the friend-making behavior as a random walk upon the LFPN naturally and captures the co-influence e?ect of the friend circles as well as personal interests to provide more accurate prediction.4. We address the continuous modeling problem of user behaviors because people’s behaviors are dynamic and the contributing factors are varying continuously over time.We present approaches to depicting the full dynamics of these factors with DP-Space,a temporal dynamic preference space, which can capture the smooth change tendency based on flexible mixtures of basis functions. Upon that we propose a generative temporal behavior model, ConTyor, in which the generation of behavior data is modeled as the joint e?ect of the dynamic social influence and varying personal preference over continuous time. We also present the EMO algorithm for model inference.5. We build the LaNES system for social network evolution analysis, providing a number of tools ranging from the evolution process inference, network structure analysis and social link prediction.